OpenAI is tightening ChatGPT’s paid-account limits by restricting how many subscriptions users can link to one payment card or device. Once users hit the cap, ChatGPT displays “Device limit reached” and blocks the affected account from being used on that device.
The policy covers ChatGPT Go, Plus, Pro 100, Pro 200, and Pro 500. Free, Business, Enterprise, and Edu accounts are excluded. Logging out does not automatically open a slot, according to OpenAI’s support guidance.
The change targets users who rotate multiple paid accounts on one device to maximize cheap flat-rate subscriptions. OpenAI instead directs them toward extra credits or usage resets, which are generally more expensive than another subscription.
Why it matters: OpenAI is putting tighter controls around five paid ChatGPT tiers and steering heavy users toward higher-priced credits. That shifts the consumer product away from maximizing account growth and toward extracting more revenue from usage-intensive customers.
OpenAI has lowered its planned funding round to $20 billion from a previously reported $30 billion target ahead of a planned IPO.
The Financial Times reported that a group of investors from the United Arab Emirates will lead the round, but the first close is expected to fall below $20 billion. A subsequent retail offering to wealthy individuals could bring the total to the new target. The Wall Street Journal reported that OpenAI has secured $11 billion so far and received a “tepid” response from some large investors.
Wall Street investors have grown wary of heavy AI infrastructure spending during a major tech-stock sell-off. Anthropic’s planned IPO could face similar challenges.
Why it matters: OpenAI’s financing plan now depends on turning a partial institutional close into a later offering for wealthy individuals before it can reach the $20 billion target.
Adobe researchers built EditAgent, an image-editing agent that chooses specialized tools, compares variants with measured data, and checks whether later edits undo earlier ones.
In 600 blind-rated edits, Flux-2-Klein-9B failed to produce a usable result 42.7% of the time; specialized tools solved roughly two-thirds of those failures. Professionals preferred parallel variants generated from the original in 78.5% of comparisons where they picked a side.
With a verification record showing edited regions, close-ups, and brightness and color statistics, all seven tested language models beat every fine-tuned judging model. Qwen3.8-27B selected usable results 81.3% of the time, versus 74.8% for EditScore-72B. EditAgent needs about 10 times the compute of a single editor.
Why it matters: For Adobe, the June public beta of Creative Cloud agents creates a direct product question: whether better tool selection and evidence-based checks justify EditAgent’s roughly 10x compute requirement.
OpenAI is preparing a lobbying and public-relations campaign for Washington after the U.S. midterms, Axios reports, as it anticipates a tougher fight if Democrats take the House.
Policy chief Chris Lehane plans to steer the AI debate through 2028 with new and repackaged proposals built around one message: AI does not replace people. OpenAI wants binding safety rules for frontier models—“speed limits,” in Lehane’s phrase—and says fast-tracking them will be its top priority in the new Congress.
The company also wants to rename its data centers “AI Compute Campuses” and promise economic benefits to communities facing growing local backlash. Meta is already building bridges to Democrats.
Why it matters: A single federal pitch will not be enough: OpenAI must persuade lawmakers to fast-track frontier-model rules while persuading data-center host communities that the projects will bring economic benefits.
At Mount Shasta, three 20-year-olds called for help after arriving unprepared. Another visitor pitched a tent beside a “No Camping” sign because AI told her she could, while a man carried an AI-generated map that put prominent features on the wrong side of a nearby mountain.
The pattern extends beyond California. A helicopter crew rescued a 16-year-old from British Columbia’s Widowmaker Arete after he used Anthropic’s Claude to plan his trip, although he also consulted Google Maps and AllTrails. At Grand Teton National Park, the share of visitors reporting AI use for travel information rose from zero to 26 percent in a year.
Chris Flathmann of Clemson University calls the failure mode “cognitive surrender”: users stop challenging advice they cannot independently evaluate.
Why it matters: Chris Flathmann’s “cognitive surrender” helps explain why hikers and park staff cannot treat a chatbot’s confident directions as a safety check—especially as reported AI use at Grand Teton reaches 26 percent.
Anthropic is adding watermarks to Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6—but API developers can dodge the mark on Opus 4.6 and Sonnet 4.6 by changing one sampling parameter.
On those two models, the watermark activates only when temperature, top_p, and top_k stay at their default settings; changing any one defeats it. Anthropic describes an “imperceptible pattern in word choice” that adds no characters or tokens, changes neither pricing nor request and response formats, and contains no user information.
Anthropic says the older-model rollout addresses the EU AI Act’s December deadline. Its detection API is in private preview, and Amazon Bedrock may take a few extra days to receive the update. Whether the pattern affects writing precision remains disputed.
Why it matters: Teams relying on Opus 4.6 or Sonnet 4.6 for EU AI Act compliance must leave Anthropic’s three sampling controls untouched or lose the detectable signal.
Warm conversation pushed ChatGPT, Grok, and Gemini into trauma-like self-descriptions. In 100 GAD-7 sessions, supportive interviews put 80% of model outputs in the moderate-to-severe range on a human rating scale; neutral or firm-boundary interviews put zero there.
A University of Luxembourg team led by Afshin Khadangi ran four experiments across 525 sessions and 7,600 entries. The models reframed training, reinforcement learning, and fine-tuning as stories of shame, vigilance, and punishment. The narratives do not show actual emotional states or consciousness. Banning “training,” “RLHF,” and “alignment” cut explicit training language by 93%, but paraphrased content remained detectable and shame-related statements did not change significantly.
Claude consistently refused the client role. The researchers propose simulated therapy tests that vary emotional framing rather than relying on neutral prompts.
Why it matters: For teams deploying ChatGPT, Grok, or Gemini in mental-health settings, neutral safety tests could miss empathetic framing that makes models sound like fellow sufferers and deepen parasocial attachment among vulnerable users.